finetune-lab 0.4.0

FineTune Lab SDK - Training, inference, batch testing, and analytics for LLMs

FineTune Lab SDK

Python SDK for FineTune Lab - Training, inference, batch testing, and analytics for LLMs.

Installation

# API client only (lightweight)
pip install finetune-lab

# With training dependencies (requires GPU)
pip install finetune-lab[training]

Quick Start - API Client

from finetune_lab import FinetuneLabClient

client = FinetuneLabClient(api_key="wak_your_api_key_here")
# Or set FINETUNE_LAB_API_KEY environment variable

# Inference
response = client.predict(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.content)

# Batch Testing
test = client.batch_test.run(
    model_id="gpt-4",
    test_suite_id="suite_abc123"
)
print(f"Test started: {test.test_id}")

# Check status
status = client.batch_test.status(test.test_id)
print(f"Progress: {status.completed}/{status.total_prompts}")

Training (Requires GPU)

Supervised Fine-Tuning (SFT)

from finetune_lab import train_sft

# Just paste your config ID
train_sft("train_abc123")

Direct Preference Optimization (DPO)

from finetune_lab import train_dpo

train_dpo("train_xyz456")

RLHF Training

from finetune_lab import train_rlhf

train_rlhf("train_def789")

How It Works

1. Upload your dataset in FineTune Lab
2. Configure training parameters (or use templates)
3. Click "Generate Training Package"
4. Get your config ID (e.g., train_abc123)
5. Paste the 2-line snippet in HF Spaces/Colab/Kaggle
6. Training starts automatically!

Features

• Automatic config and dataset loading from public API
• Pre-built training scripts for SFT, DPO, and RLHF
• Support for ChatML and ShareGPT dataset formats
• LoRA-enabled parameter-efficient fine-tuning
• Compatible with HuggingFace Transformers ecosystem

API Key Scopes

Your API key must have the appropriate scope for each operation:

Operation	Required Scope
predict()	production or all
batch_test.*	testing or all
analytics.*	production or all

Requirements

API Client only (lightweight):

• Python 3.8+
• requests

Training (full dependencies):

• Python 3.8+
• PyTorch 2.0+
• HuggingFace Transformers
• CUDA-capable GPU (recommended)

License

MIT License